Informed voters and electoral outcomes: a natural experiment stemming from a fundamental information technological shift

Page created by Evelyn Sherman
 
CONTINUE READING
Informed voters and electoral outcomes: a natural experiment stemming from a fundamental information technological shift
Public Choice
https://doi.org/10.1007/s11127-021-00884-z

Informed voters and electoral outcomes: a natural
experiment stemming from a fundamental
information‑technological shift

Shane Sanders1 · Joel Potter2 · Justin Ehrlich3 · Justin Perline4 ·
Christopher Boudreaux5

Received: 6 February 2020 / Accepted: 27 January 2021
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC part of Springer Nature 2021

Abstract
Do informed electorates choose better candidates? While that question is straightforward,
its answer often is elusive. Typically, candidate-quality information is neither salient nor
subject to exogenous change. We identify a natural experiment within a non-political
election setting that is transparent and features exogenous change in the candidate-qual-
ity information frontier. The setting is Major League Baseball’s (MLB) annual selection
of two most valuable players, a challenging environment with an innately heterogeneous
candidate set, and the exogenous change is the development of the pathbreaking, compre-
hensive player-value measure Wins Above Replacement (WAR​) in 2004 and its subsequent
calculation for all retrospective MLB player-seasons. WAR​’s development and rapid popu-
larization informed voting from 2004 onward. Retrospective calculation allows us to draw
back the curtain and evaluate how pre-2004 voters behaved with respect to revealed can-
didate quality. From negative binomial, fixed-effect regression models, we find robust evi-
dence of significant, substantial, pivotal behavioral change on the part of voters since 2004.

Keywords Voter information · Voting behavior · Candidate quality · Information frontier
shift · Wins above replacement · Major league baseball

JEL Classification D72 · D80 · Z29

* Shane Sanders
 sdsander@syr.edu
1
 Syracuse University, 301 MacNaughton Hall, #316, Syracuse, NY 13244, USA
2
 University of North Georgia, 105 Barnes Hall, Dahlonega, GA 30597, USA
3
 Syracuse University, 300 MacNaughton Hall, #320, Syracuse, NY 13244, USA
4
 Pittsburgh Pirates, Pittsburgh, PA, USA
5
 Florida Atlantic University, 777 Glades Road, Kaye Hall 145, Boca Raton, FL 33431, USA

 13
 Vol.:(0123456789)
Public Choice

1 Introduction

Do informed electorates choose better candidates? While that question is straightforward,
its answer typically is elusive in large political elections. Under normal circumstances,
the level and quality of information possessed by voters is neither salient nor subject to
exogenous change. Herein, we identify a natural experiment within a non-political election
setting that is transparent and features an exogenous change in the technology or frontier
of candidate quality information. The setting is voting in the selection of Major League
Baseball’s American League and National League Most Valuable Players (MVPs), and the
exogenous change is the development of the comprehensive player value measure Wins
Above Replacement (WAR​). The WAR metric represented an exogenous change because
the measure was not created for the purpose of electing MLB’s two MVPs. Rather, it was
created during the 2004 offseason for the de facto purpose of informing popular “sporto-
metric” (Goff and Tollison 1990) research on baseball. However, it had a tertiary effect of
making candidate information salient.
 Despite its non-political setting, the present study focuses on a transparent election of
substantial financial significance.1 In that environment, we consider the role of informa-
tion in addressing the question of whether informed electorates choose better candidates
(Banerjee et al. 2011; Pande 2011). The study’s setting contains several generalizable ele-
ments. For example, we are able to consider voters’ response to changes in overall can-
didate quality and also their response to changes in constituents of candidate quality. We
also consider the extent to which voters respond to non-meritorious candidate factors. For
example, it may be that voters respond to information about candidates’ races, ages, or
teammate quality, even conditional on the knowledge of specific candidate quality. In the
present setting, MVP candidates produce baseball value in a team setting. Therefore, we
also consider whether voters respond to teammate quality, conditional on the candidate’s
own quality. That possibility has analogs to political elections in that political candidates
typically campaign as members of a political party. Each of those considerations is exam-
ined empirically both in general and also conditional on changes in the candidate quality
informational frontier that forms the paper’s natural experimental setting. Through its focal
points, the study seeks to understand generalizable aspects as to the behavior of newly
informed voters.
 The present study represents a coordination problem and not so much a problem of con-
vergence to true player value.2 In particular, we consider whether voters coordinated on
different salient proxies for player value before the creation of Wins Above Replacement
(WAR​) than afterwards.3,4 WAR​ provides a refined estimate to answer the following ques-
tion (Slowinski 2010): “If this player got injured and [his] team had to replace [him] with

1
 Some teams adopt MVP awards as an incentive mechanism. Some players are offered large contractual
bonuses for winning being selected as an MVP. For example, star MLB player Mike Trout earns an addi-
tional $500,000 by his team, the Los Angeles Angels, in the event of winning an MVP award.
2
 We thank an anonymous editor for pointing this out.
3
 The formula for position players is WAR = (Batting Runs + Base Running Runs + Fielding Runs + Defen-
sive Runs Saved + Position Adjustment + Replacement Runs)/(Average Runs Needed to Obtain an Addi-
tional Win).
4
 The formula for pitchers is WAR = [(RA9avg—RA9)*(IP/9) + Rlr]/(Rpw) where RA9avg is the number
of runs an average pitcher is expected to give up in 9 innings, RA9 is the average number of runs given up
per nine innings, IP is the number of innings pitched, Rlr is replacement level runs, and Rpw is the average
number of runs needed to obtain an additional win.

13
Public Choice

a freely available minor league player, how much value would the team be losing?”. In the
case of a player with a WAR​ value of 5, replacing that player with a minor league player
would result in an estimated five games lost to the team, ceteris paribus. As such, WAR​ is
designed to measure a player’s performance independent of his teammates.5 Following the
development of WAR​, the measure may represent only one of multiple inputs that voters
consider in selecting a most valuable player. For example, some voters may consider both
WAR​and other performance measures, such as a player’s win pivotality in close games or a
player’s win pivotality in important games, when estimating player value.

1.1 MLB’s election(s): a brief institutional background

Major League Baseball was formed as a confederacy of two leagues—the National League
(NL; 1876-) and the American League (AL; 1901-)—in 1903. Since 1911, the NL and AL
each have chosen distinct Most Valuable Players following every regular season. Begin-
ning in 1931, the MVP awards have been selected by the Baseball Writers’ Association of
America (BBWAA​). The Baseball Almanac (2019) summarizes the award’s early history:
 There have been three different official Most Valuable Player awards in Major
 League Baseball history, since 1911; the Chalmers Award (1911–1914), the League
 Award (1922–1929), and the Major League Baseball Most Valuable Player Award
 [1931-]. The … MVP is presented annually by the BBWAA​. It is considered by MLB
 as the only official Most Valuable Player Award and symbolizes the pinnacle of a
 player’s personal achievement during any single season of play.
 In 1938, the BBWAA​began electing MVPs on the basis of voting by its members. Initially,
three NL (AL) award voters were designated for each NL (AL) team city. That number was
reduced to two in 1961. For several years, then, selections have been made by 60 MLB
voters, 30 participants for each league’s election. The voting rule employed is a weighted,
truncated Borda method of marks that has also been called a corner-weighted version of
Borda’s rule (see, e.g., Greenwich et al. 2019). Studies have examined the weighted, trun-
cated Borda method with applications to Formula One racing and ski jumping competi-
tions (Kaiser 2019). When the Borda count is applied to elections, all voters rank candi-
dates or policy alternatives from most preferred to least preferred. Typically, the lowest
ranked alternative is assigned 1 point, the next lowest 2 points, and so on until the highest
ranked alternative receives n points (Kaiser 2019). Here, n is the number of candidates
involved in an MVP election. Points are tallied for each candidate, and the candidate with
the largest total wins.
 In MLB, players receive MVP votes much like candidates in a political election. How-
ever, the voting procedure adopted for selecting a winner is a weighted truncated Borda
count rather than the classic Borda method of marks described above.6 It is weighted
because the point intervals between different consecutive pairs of ranked options are not
the same; it is truncated because only a subset of all alternatives is evaluated (i.e., other,
lower ranked candidates receive zero marks). Each voter ranks 10 players who earn points

5
 A teammate effect is, however, possible. For instance, if a good player has an even better player hitting
behind him in the batting order, it is plausible the first batter will receive better pitches to hit and would
therefore produce more runs and wins (which is not captured by WAR​).
6
 Young (1995) discusses Borda voting and its implications for the likelihood of identifying the “correct”
(Condorcet) top candidate should one exist.

 13
Public Choice

according to the following allocation: (14, 9, 8, 7, 6, 5, 4, 3, 2, 1). The points for all players
are summed to determine the final score. Under that system, the player-candidate with the
highest vote-point total for a given league-year is crowned league-year MVP. Under MLB’s
procedure, a total of (14 + i=1 i)× 30 = 1770 points are available for a given league’s race
 ∑9
and one player can receive as many as (30 × 14) = 420 points.7

1.2 Informed voting and the development of WAR: a natural experiment?

Baseball is a game of specialization such that cross-positional value comparisons have an
apples-to-oranges quality, especially in the absence of a comprehensive, cross-positional
measure of contributions to winning games. Consider a comparison between a start-
ing pitcher and a shortstop. The pitcher attempts to throw the ball past the batter and either
strike the batter out or force an out with the help of his defending teammates. Hence,
the pitcher’s position is almost purely defensive. However, the shortstop (i.e., the player
between 2nd and 3rd base) plays substantial roles on defense (i.e., catching the ball and
stopping the opposing player from advancing) and offense (i.e., hitting the ball). The short-
stop accumulates win value by mixing his offensive and defensive skills, but the pitcher
accumulates value almost exclusively by throwing baseballs and does so by taking the
pitching mound about once every six days. Without adopting some exchange rate mecha-
nism, it is difficult to compare the two players’ respective contributions. And yet, MVP
voting demanded just such a comparison for decades. In 2004, WAR​ was created by Base-
ball Prospectus writer Jay Jaffe (2004) as a cross-positional measure of player win value
(above league replacement level value at player position). From 2004 on, WAR​ has been
calculated for all MLB player-seasons by Baseball Prospectus and other baseball publi-
cations. What is most important, it also has been calculated retrospectively for the uni-
verse of professional baseball. Baseball Reference (www.baseb​all-refer​ence.com) features
retrospective calculations beginning with the first season of the first professional baseball
league (National Association of Professional Baseball Players, 1871–1875) and continuing
through the respective histories of the NL (1876-) and AL (1901-). Those histories were
archived using the massive, crowd-sourced data collection project initiated by retrosheet.
org, which was created by University of Delaware biology professor David Smith in 1989
(Smith 2000).
 The concurrent assembling of past and present WAR​ data since 2004 has created some-
thing of a natural experimental setting. Since 2004, MLB’s MVP voters have cast ballots
largely with knowledge of player WAR​ values. Before 2004, voters were comparatively
uninformed. Despite that lack of information, retrospective WAR​ values subsequently
became available for past player seasons such that we can evaluate both earlier and later
MVP voting behavior with respect to WAR​ values. Given WAR​’s prominence (on, e.g.,
baseballprospectus.com, fangraphs.com, and baseball-reference.com) and the institutional
nature of the BBWAA​ as a defined group with a voluminous archived website, as well as
regular chapter and national meetings, Baseball Writers who now vote for MLB Awards
every year presumably are better informed voters than their pre-2004 counterparts.8

7
 In our dataset, the NL and AL were imbalanced in terms of numbers of teams from 1998 (when the Mil-
waukee Brewers moved to the NL) through 2012 (after which the Houston Astros moved to the AL). That
imbalance affected the number of ballots cast for each league during those years. We account for it empiri-
cally by re-scaling ballots to a maximum of 420 possible points for each player during those seasons.
8
 Many of the counterparts are earlier versions of the same person.

13
Public Choice

Baseball Writers refer frequently to WAR​ (see, e.g., Madden 2017, NY Daily News and
Slowinski 2010 for summaries of the measure’s importance to Baseball Writers).

1.3 Literature and motivation

The present study takes advantage of the development and retrospective calculation of
WAR​as a natural experiment by which to ask a question that Banerjee et al. (2011) consid-
ered previously in a starkly different voting context. Specifically, we ask whether informed
voters make better choices in the presence of objective information about candidate qual-
ity. Whereas fans, participants, and voters may not want the MVP race to boil down to a
contest of highest WAR​, neither are they likely to wish for the race to be too far removed
from considerations of objective player values. While WAR​ is the central measure of play-
ers’ contributions to winning baseball games, voters also may consider alternative dimen-
sions of player win value, such as the player’s specific contribution toward winning in close
or important games. WAR​ considers how a player’s productivity relates to team winning
given average game settings and therefore does not weight player performances in win-
ning pivotal game scenarios. Banerjee et al. (2011) conducted field voting experiments in
India and found evidence that non-partisan, third-party public disclosure of incumbent leg-
islator report cards led to larger vote shares for high-performing incumbents. Within the
present voting context, the WAR​ measure can be thought of as similar to a non-partisan,
third-party evaluation of candidates. It evaluates candidates in a way that imposes no a
priori subjective criteria and is “computationally agnostic” to the contest’s results. Pande
(2011) reviews the literature and concludes the findings of Banerjee et al. (2011) to be
robust across different voting environments. Baron (1994) constructs a model of electoral
competition with informed and uninformed voters. He finds that informed voters have the
effect of reducing the legislative influence of interest groups in equilibrium as well as the
overall level of campaign financing by interest groups.
 Pietryka and DeBats (2017) examine the relationship between voting behavior—includ-
ing voter information—and proximity to elites in one’s social network. Ashworth and De
Mesquita (2014) demonstrate within a set of formal models that the relationship between
voter information and democratic performance may be positive or negative depending on
a set of parametric values. Thoth and Chytilek (2018) find that voters facing time pres-
sure shift their information-gathering efforts from accuracy to efficiency when evaluating
candidates. Specifically, voters restrict their attentions to a smaller set of policies in those
evaluations. MLB’s voters may face time pressures in that the performance differences of a
season’s top players often are slight and subtle and may require viewing hundreds of hours
of game footage to perceive. In a given season, a player could be on the roster of any of the
30 MLB teams, each of which plays a 162-game schedule. MLB schedules 2,480 games in
a season and the average game runs a little over three hours, according to Baseball Refer-
ence. As such, approximately 7,300 h—more than 304 24-h days—of regular season games
are played every year. Faced with that massive quantity of game performances, time-
friendly measures to evaluate player-candidates likely are necessary for voters. Several
other contributions to the relevant literature study alternative dimensions of information
and voting (see, e.g., McMurray 2015; Jensen et al. 2015; Garmann 2017; Ginzburg 2017).
 MLB and sports leagues in general are well-established in the literature as laboratories
for public choice, economics of regulation, and labor economics study (see, e.g., Shughart
1997, antitrust legislation related to baseball; Ross and Dunn 2007, income tax responsive-
ness of MLB All-Stars; Goff et al. 1998, moral hazard in baseball; Hill and Groothius 2001,

 13
Public Choice

redistribution of rents among teams; Ehrlich et al. 2020, labor policies during pandemics).9
As both “America’s Pastime” and a sport that lends itself to statistical analysis, MLB pro-
vides a long history of labor productivity data that is unparalleled in many respects within
American industrial enterprises and even among other popular American sports.
 The remainder of the paper is structured as follows. Section 2 presents the dataset used
in our study. Controlling for a set of team and season fixed effects negative binomial vote-
count regression models,10 Sect. 3 specifies and tests a model of informed voting. Specifi-
cally, the model asks whether MLB elections from 2004 relate more strongly to an objec-
tive, comprehensive, cross-positional measure of on-field value (i.e., WAR​) than do prior
elections dating from 1980. That is, are informed voters making choices that allow them to
identify more accurately the true “Most Valuable Player” (rank-ordering) for each league-
season? In the absence of information on players’ WAR​, we consider the factors that may
have been more salient for evaluating MVP candidates before 2004. Section 4 presents and
discusses the study’s central results. Section 5 concludes.

1.4 Data description, summary and visualization

We collected data on MLB’s top 50 seasonal WAR​ leaders for every regular season from
1980 to 2017. The sample comprises 38 vote-years and 76 league-level elections. We rely
on the Baseball Reference version of WAR​for our study rather than the Fangraphs version,
the Baseball Prospectus version, the openWAR​ version, or another implementation. While
all WAR​ measures generate positively correlated sets of WAR​ values of moderate to high
strength (see, e.g., Baumer et al. 2015), slight methodological differences exist across the
implementations; those differences are beyond the scope of the present study.11 We chose
Baseball Reference WAR​because it is popular, accessible, and was created with retrospec-
tive analysis in mind.12 Allowing for possible ties in seasonal WAR​ value (at the fiftieth-
highest value), the sample contains 1907 player-season observations rather than 1900. Each
MLB voter is charged with identifying and ordering ten Most Valuable Players in a given
league; Baseball Writers typically exhibit strong consensus as to whom should be on a
given ballot. For example, the union of all AL ballots in the year 2000—approximately
midway through our sample—contained 19 AL players and 22 NL players.
 The present dataset was retrieved, with permission, from baseball-reference.com
using the RVest package in the statistical software program R. The primary variables at
the player-season level include13: (election) vote point count, WAR value, age, whether
on a team in a big (top-5) market, whether traded during season, (primary) league of
play, (primary) playing position and (primary) team-of-play. Summary statistics for

9
 More generally, sport often is used to measure the empirical incidence of aggregation paradoxes. For a
recent contribution along those lines, see, e.g., Boudreau et al. (2018).
10
 Negative binomial models are chosen over Poisson models owing to strong evidence of overdispersion in
vote-count data.
11
 All WAR measures use the same inputs but slightly different specifications.
12
 Specifically, the measure was designed for description rather than for forecasting. Baseball Reference
Founder and CEO Sean Forman confirmed that point in a conversation with two of the present authors. Of
course, retrospective analysis of the recent season is an input into MVP decision-making.
13
 We do not control for game characteristics because it is unlikely that they will affect an individual’s
WAR score and MVP votes. The reason is that because if a player is a season’s WAR leader, he always has
participated in a large number of games (n > 100). Over the course of a season, we therefore expect game
characteristics to revert to the mean against an assortment of different opponents.

13
Public Choice

Table 1  Summary statistics of Variable Obs Mean St dev (s) Min Max
key variables
 Votepts 1907 54.33 92.68 0 420
 WAR​ 1907 5.60 1.29 3 12.7
 Age 1907 28.30 3.69 19 44
 Big market 1907 0.30 0.46 0 1
 Multiple teams 1907 0.02 0.15 0 1
 American league (AL) 1907 0.50 0.50 0 1
 Pitcher (P) 1907 0.31 0.46 0 1
 Catcher (C) 1907 0.05 0.21 0 1
 First base (1B) 1907 0.10 0.30 0 1
 Second base (2B) 1907 0.07 0.26 0 1
 Shortstop (SS) 1907 0.07 0.26 0 1
 Third base (3B) 1907 0.11 0.31 0 1
 Right fielder (RF) 1907 0.09 0.29 0 1
 Center fielder (CF) 1907 0.11 0.31 0 1
 Left fielder (LF) 1907 0.08 0.27 0 1
 Designated hitter (DH) 1907 0.01 0.12 0 1

these variables are provided in Table 1. Note that vote point count is rescaled from 1998
through 2012 to account for league imbalance.
 Player age in the sample ranges between 19 and 44, which is consistent with most
of the overall age range for Major League Baseball. Incredibly, the average WAR​ value
in the sample is 5.60. That value suggests that MLB stars generate exceptional average
win value. A team of replacement players commonly is estimated to win about 48 games
in a season, ceteris paribus (Baseball Reference 2012). Using that benchmark, we can
consider the implied success of a (25-player) MLB team in which players average a WAR​
value of 5.60. In fact, such a high level of win production could not exist on one team
(owing to crowding out effects). A single team would have to win 48 + (5.60 × 25) = 188
games to support such a WAR​ average. However, only 162 MLB games are played in
regular season. The sample is balanced in terms of league representation. In the dataset,
playing position is represented by a set of position dummy variables, {P, C, 1B, …,
DH}. We observe substantial variation in sample representation by playing position, an
observation that stands to reason because (players at) one position innately are more
valuable than (players at) others. We observe that more than 28% of sampled player-
seasons were conducted in the service of a team in a city of top-5 market size (repre-
sented as the variable big market). Eight of 30 (26.7% of) MLB teams were located in
a top-5 market throughout the sample; it therefore appears that top baseball talent does
not gravitate disproportionately to MLB’s biggest markets despite the absence of a sal-
ary cap in MLB. We base our market-size variable on the US metropolitan areas with
the five largest populations according to the 1980, 1990, 2000, and 2010 US Censes.
Surprisingly, the five metropolitan areas on that list did not change from 1980 to 1990,
to 2000, or to 2010. In the estimation section, we will consider the distinct question as
to whether being in a “big market” helps a player in terms of MVP voting points. As
that variable is related to team fixed effects, the two will be treated as empirical sub-
stitutes in that section. Lastly, we observe that players represented multiple teams in 43
player-season pairs.

 13
Public Choice

Fig. 1  Kernel density plot of vote points

 The sample contains player-season level vote counts, but not voter-level (ballot-level)
observations. Ballot-level data were not disclosed by the BBWAA​ until 2018. As such, we
treat aggregate player vote point counts as our main left-hand side variable. While voter
characteristics largely are unavailable to the public, we note that all active members of the
BBWAA​are eligible to vote. As such, the base of voters for the MVP award is quite persis-
tent over time. To verify that conclusion, we examined a subset of individual ballots that
have been made public since 2012 (but unfortunately not before the creation of WAR​) at
https​://bbwaa​.com/voter​-datab​ase/. Using single letter search queries, we sampled the user
database randomly to identify 40 BBWAA​members who voted for the MVP awards at least
once during the observation period. Of those 40 members, 35 had voted in multiple elec-
tions during the period of our sample. While we cannot control for individual voter char-
acteristics, the observed persistence in the set of voters, which is an institutional character-
istic, suggests that voter characteristics also are persistent across time. We take advantage
of a long longitudinal dataset with substantial variation in player characteristics. We also
benefit from the natural experimental context of WAR​’s development (with subsequent ret-
rospective calculations). From our 1907 player-seasons, 1125 candidates for MVP received
at least one vote point. With complete consensus across ballots (as to whom should receive
a vote), 760 players would have received points. With no consensus (with respect to play-
ers in the dataset), all 1907 sampled player-seasons would have received votes. Thus, we
observe that Baseball Writers exhibit moderately strong consensus as to whom should be
on the ballot(s). The present study considers whether such a level of consensus was formed
among informed voters or formed among uninformed voters (e.g., an echo chamber) over
time. Figure 1 displays a kernel density plot of the vote point count within our sample.
 Figure 1 depicts a smoothed representation of a discrete variable (vote point count).
The figure suggests that, even among seasonal WAR​ leaders, most player-seasons result
in a small number of vote points. The median vote point count in the sample is 4, and
the 75th percentile is 64 points. Given our model’s specification (i.e., of a count data
model), we account for the extreme right skewness of the dependent variable, its discrete

13
Public Choice

Fig. 2  Distribution of possible player vote point counts

distribution, and its left truncation (at zero). Vote point count is a count variable in that (a)
it is non-negative integer valued, (b) it in fact tabulates how many points voters allocate
to a given player (from their bundle of points), and (c) the possible points allocated to a
player represent a set of consecutive, non-negative integers along with one possible non-
consecutive
 { value obtained by a player sweeping all first place votes. That is, Vote Point
Counti = vi ∈ N ∶ 0 ≤ vi ≤ 415 ∪ v = 420 . To demonstrate that vote point count occu-
 }

pies that co-domain, Fig. 2 displays a histogram of the distribution of possible vote point
counts for an individual voter (i.e., from the set of all possible 30 × 1 voting vectors for a
given voting individual), where the frequency count on the y-axis represents the number of
possible vectors yielding the corresponding number of vote points.
 Before formally modeling the relationship of interest, let us consider a scatter plot with
player-season vote point count on the vertical axis and player-season WAR​ value on the
horizontal axis. In Fig. 3a, we adopt color-coded data points and corresponding color-
coded trend lines to depict the uncontrolled relationship between the two variables both
before and after the creation of WAR​, respectively.
 The uncontrolled relationship between vote point count and WAR​was positive both from
1980 through 2003 and from 2004 through 2017. However, simple regression trend lines
suggest that the responsiveness of vote point count to changes in WAR​noticeably has been
stronger in the latter period. Multivariate regression analysis will inform us as to whether
this apparent difference is significant and substantial conditional on a set of control vari-
ables. When considering Fig. 3a, one might wonder whether the apparent slope difference
is indeed explained by the creation of WAR​ or if it might suggest a more general (gradual)
improvement in voters’ knowledge and ability over the course of the dataset. We are deal-
ing herein with panel data having multiple observations per period. Therefore, an endog-
enous structural break estimation is not supported. However, we do in fact know when the
information technology breakpoint of interest occurred (a few months after the 2003 vote),
such that an exogenous structural break estimation is appropriate. Moreover, we know that

 13
Public Choice

Fig. 3  a Plot of MVP vote points against WAR for 2 time periods (before and after information technologi-
cal shift). b Plot of MVP vote points against WAR for 3 time periods (before information technological shift
in 2 periods and then after)

WAR​ is the first and only virally adopted cross-positional player value measure (i.e., the
first such measure to be featured on leading baseball statistics sites Baseball Prospectus,
Baseball Reference, ESPN, and Fan Graphs). As such, WAR​was not established in a sea of
comparable performance measures. Rather, it was a groundbreaking measure in the area of
comprehensive sabermetric player analysis. Baumer and Matthews (2014) state in an arti-
cle subtitled There is No Avoiding WAR​:

13
Public Choice

 While there have been many important contributions to the field, arguably the most
 prominent success story in recent years has been wins above replacement (WAR​).
 WAR​ is an all-encompassing assessment of a baseball player’s contribution to his
 team, measured in wins added relative to a hypothetical (and often vaguely defined)
 “replacement player.” ….While predecessors of WAR​ (like Bill James’s Win Shares)
 have been around for some time, the modern incarnation of WAR​ may now reach a
 broader audience than any sabermetric stat since OBP. Perhaps the most telling indi-
 cation of WAR​’s permeation of the baseball landscape was the announcement that
 WAR​will appear on the back of Topps baseball cards in 2013.
 According to Baumer and Matthews (2014), the only sabermetric measure comparable
to WAR​ in impact is On Base Percentage, which measures a player’s value in one aspect
of the game. For the present setting, then, no other technological breakpoint candidates
exist other than 2004. Therefore, we will conduct a Chow (exogenous structural break)
Test—through our empirical specification itself—in the estimation section to follow. The
structural break testing initially will center on the year 2004 before considering the pos-
sibility of an earlier breakpoint. Despite not having a tractable endogenous structural break
test at our disposal, we will rely on one means of testing for a more gradual change in
voter behavior. In the visualization to follow, as well as in the estimation section, we divide
our pre-2004 sample into two equal sub-samples: 1980–1991 and 1992–2003. If voters
were simply improving in their ability to measure player value gradually, we might expect
the slope of the trend line between vote point count and WAR​ to increase incrementally
between both the first (1980–1991) and second (1992–2003) sub-samples and between the
second and third (2004–2017) sub-samples. The trinary color-coded plot in Fig. 3b helps
us consider that relationship across time.
 In terms of the (slope of) relationship between vote points and WAR​, no substantial dif-
ference is evident between the first two periods. As such, Fig. 3b is quite similar to Fig. 3a
as far as identifying a potential change in the relationship of interest beginning in 2004
goes. To examine further if 2004 represents a structural breakpoint, we plot the mean sea-
son-league Spearman rank correlations between player WAR​ and player MVP vote points
for each of the three periods in Fig. 4.
 The time series plot of Spearman rank correlation coefficients shows little difference in
means between the first two periods. However, the sample mean increased markedly during
the period from 2004 onward. In our estimation models, we will consider both binary and
trinary partitions of the dataset within a multivariate regression setting.

2 Model specification

Given that the dependent variable, vote point count, is a non-negative integer-valued varia-
ble, we specify a count model herein. For all but one model, we forego the specification of a
traditional linear model in favor of a count model because count data are zero-censored and
typically right-skewed. The observations likewise are sparse because the variable takes on
only integer values, and that sparseness contributes to non-normally distributed residuals if
one explains variation in a count dependent variable by estimating a traditional linear model
(Cameron and Trivedi 2013). A chi-squared test reveals over-dispersion in our dependent vari-
able such that we adopt a negative binomial count model rather than a Poisson count model.
The baseline model contains team, season, and playing position as fixed effect variables; a
Hausman specification test supports selection of fixed over random effects modeling for each

 13
Public Choice

Fig. 4  Mean spearman rank correlation between WAR and MVP votes by time period

independent variable. Given the institutional and empirical setting of the dataset, we forego
entering player fixed effects. Many sampled players (283) appear in the dataset in one season
only. Several other players (70) appear in the dataset in more than one season but do not earn
an MVP vote in any season. Moreover, team and position fixed effects are mutually compatible
within one specification, whereas the specification of player fixed effects precludes entering
both position and team, as explanatory variables because players tend to remain on the same
team’s roster and play the same position. Thus, specification of player fixed effects would lead
to (a) loss of substantial data representing specific types of player-seasons (sub-sample selec-
tion bias) and (b) preclusion of our other fixed effect variables. As such, we decide that these
data do not have classic micro-panel characteristics and specify a panel structure at the levels
of team, position, and season.
 Our set of fixed effect, negative binomial regressions appear as follows. The baseline model
tests for a general relationship between vote point count and WAR​ from 1980 to 2017. We
specify one model with team fixed effects and one model with big market in place of team
fixed effects. We do so because big market is collinear with team fixed effects and thus cannot
enter the same model. The baseline models are specified as follows.

 voteptsi,t = 0 + 1 WARi,t + 2 agei,t + 3 age2i,t + 4 multiplet eamsi,t
 (1)
 + 5 i + 6 i + 7 t + i,t

 voteptsi,t = 0 + 1 WARi,t + 2 agei,t + 3 age2i,t + 4 multiple_teamsi,t
 (2)
 + 5 bigmarketi + 6 i + 7 t + i,t

13
Public Choice

 Nothing ensures a priori that vote point count and WAR​ bear a significant relationship
over the full data sample. The joint purpose of the baseline models is to test whether any
such relationship exists. As WAR​was not created until 48 of the 76 sampled votes had been
cast, voters made decisions in the absence of WAR​ for almost two-thirds of the sample
period. Over that timeframe, voters cast ballots presumably based on their observations of
the game and on available (decentralized) player statistics. Therefore, no a priori assurance
exists that voters adopted WAR​in their voting behavior following its creation. While Base-
ball Writers certainly have featured WAR​ prominently in (public) baseball articles since
2004, it is an empirical question as to whether that adoption carried over to their (individu-
ally private or anonymous) voting behavior.
 Following the baseline treatment, we adopt a Chow Test (for difference in slopes) to test
for differences in the relationship of interest during the 2004–2017 sub-period. Like the
scatter plots presented in the previous section, the same test is conducted with 1980–2003
as the (single) comparison period and (in subsequent specifications) with 1980–1991 and
1992–2003 as the (dual) comparison periods. The intercept in our models represents the
expected vote point count of a replacement player-season (i.e., a player with a WAR​ value
of 0). Such player-seasons are well below average by definition and thus not observed in
the data. Our two-period Chow Test specifications will be conducted as follows.

 voteptsi,t = 0 + 1 WARi,t + 2 WARi,t xafter_2003i,t + 3 agei,t + 4 age2i,t
 ( )
 (3)
 + 5 multiplet eamsi,t + 6 i + 7 i + 8 t + i,t

 voteptsi,t = 0 + 1 WARi,t + 2 WARi,t xafter_2003i,t + 3 agei,t + 4 age2i,t + 5 multiple_teamsi,t
 ( )

 + 6 big_marketi,t + 7 i + 8 t + i,t
 (4)
 The variable, after_2003, is an indicator that is set equal to 1 for MVP voting years
2004–2017 and 0 otherwise. We interact that variable with WAR​ to compare the relation-
ship between vote point count and WAR​ before and after the measure’s creation. The vari-
able after_2003 does not enter the right-hand side of the model alone for reasons discussed
previously, namely, that no reason exists for expecting a difference in intercept values
across time periods. Moreover, the intercept is extrapolative within the model, as replace-
ment players are not expected to receive votes. We next consider a model that is consistent
with a three-period Chow Test for changes in slope.
 ( )
 voteptsi,t = 0 + 1 WARi,t + 2 WARi,t xafter2003i,t + 3 WARi,t xbetween_92&03it + 4 agei,t + 5 age2i,t
 ( )

 + 6 multiple_teamsi,t + 7 i + 8 i + 9 t + i,t
 (5)
 ( ) ( )
 voteptsi,t = 0 + 1 WARi,t + 2 WARi,t xafter_2003i,t _ 3 WARi,t xbetween_92&03it
 + 4 agei,t + 5 age2i,t + 6 multiple_teamsi,t + 7 big_marketi,t + 8 i + 9 t + i,t
 (6)
 The variable between_92&03 allows us to trisect the data so as to test for possible evi-
dence that voting behavior actually began to change before 2004 (e.g., as some more grad-
ual result of the sabermetric movement in general). In essence, models (5) and (6) allow
us to conduct a Chow Test for changes in the relationship between vote points and WAR​for
the periods 1980–1991, 1992–2003 and 2004–2017. If the creation of WAR​ were integral

 13
Public Choice

to changing that relationship, we might expect it to remain fairly stable before 2004 and to
change significantly (and perhaps substantially) thereafter.

3 Estimation and results

 In this section, we report the estimation results for the six fixed effect, negative binomial
 models specified in the previous section. These results are presented in Table 2.
 Over the full sample, Table 2 demonstrates that the relationship between WAR​and MVP
 vote points is positive and significant. Models 3 and 4 show that the estimated slope of
 that relationship increased significantly (became significantly more positive) following the
 creation and publication of WAR​. Models 5 and 6 provide evidence that the increase did
 not arise as a gradual and vague response to the sabermetric era in general (e.g., not as a
 process that began in years prior to 2004 and built from there). Rather, the models demon-
 strate that voters behaved in a manner from 1992 to 2003 statistically equivalent to their
 behavior from 1980 to 1991. It is only in the 2004–2017 sub-sample that we observe a
 change in voter behavior. As such, we have evidence that (informed) post-2003 MVP vot-
 ers allocated vote points in ways that are more consistent with actual player value than did
 their relatively uninformed counterparts of earlier periods. Furthermore, the explanatory
 power of the model (i.e., in terms of improved ­R2) rises following the creation of WAR​.
 For the purpose of assessing explanatory power before and after the development of
WAR​, we divide our sample into two sub-periods (1980–2003 and 2004–2017). We then
estimate models (1) for each period separately as ordinary least squares, fixed effects
regressions. We do so because fixed effects, negative binomial estimation does not yield
­R2 coefficients or any other intuitive measure of explanatory power. We find that overall
 ­R2 rises from 0.366 in the former sample to 0.464 in the latter. That is, the percentage of
 variation in vote points explained by variations in player characteristics and performances
 rose by almost 10 percentage points in the latter period. It is not simply that voting was
 significantly and substantially more responsive to estimated player contributions to win-
 ning games beginning in 2004. From 2004 forward, voting has been explained better by
 measures of player value (i.e., has been less subject to noise).
 Negative binomial regression coefficient estimates can be interpreted as log-linear semi-
 elasticities. For example, model (3) suggests that each additional unit of WAR​ increased
 players’ expected vote point count by 52.8% prior to 2004 and by 52.8 + 7.8 or 60.6% after
 2004. Models (4)–(6) provide similar semi-elasticity results. If pre-2004 voting had been
 as responsive to WAR​ as subsequent voting, ceteris paribus, several pre-2004 races would
 have pivoted (e.g., the 1999 and 2001 AL races). Furthermore, we replace WAR​with offen-
 sive WAR​and defensive WAR​for position players and with pitching WAR​or pWAR​ for
 pitchers. We estimate a modified version of model (3) so as to determine what specific
 aspects of performance were treated differently by voters after 2004. We run the modified
 version as a negative binomial model for both (1) position players and for (2) pitchers.
 Although we estimate the full specification of Table 2’s Model (3), we report only coef-
 ficient estimates for the variables of interest in Table 3 (for brevity).
 The results of Table 3 suggest that voters responded to all forms of change in candi-
 date-quality: offensive WAR​, defensive WAR​, and pitching WAR​. For each constituent value
 measure, a significant increase in the corresponding coefficient from 2004 is observed.
 Before the advent of WAR​, no statistical evidence is found that ceteris paribus improve-
 ments in defensive performance (defensive WAR​or dWAR​) raised one’s MVP vote point

13
Public Choice

Table 2  Main estimation results
 (1) (2) (3) (4) (5) (6)
 VotePts VotePts VotePts VotePts VotePts VotePts

WAR​ 0.545**** 0.529**** 0.528**** 0.514**** 0.544**** 0.532****
 (0.018) (0.017) (0.018) (0.017) (0.020) (0.020)
Age − 0.004 0.052 0.027 0.113 0.034 0.116
 (0.089) (0.086) (0.088) (0.085) (0.087) (0.084)
Age2 0.000 − 0.001 − 0.000 − 0.002 − 0.001 − 0.002
 (0.002) (0.001) (0.001) (0.001) (0.001) (0.001)
Multiple teams − 0.791*** − 0.849**** − 0.849****
 (0.257) (0.258) (0.258)
Bigmarket 0.039 0.045 0.053
 (0.061) (0.060) (0.060)
War ×after03 0.078**** 0.077**** 0.065**** 0.063****
 (0.012) (0.012) (0.015) (0.014)
War × between − 0.024 − 0.027*
 _92&03
 (0.016) (0.015)
First base (1B) 0.795**** 0.817**** 0.758**** 0.780**** 0.776**** 0.808****
 (0.120) (0.116) (0.119) (0.116) (0.120) (0.117)
Second base (2B) − 0.081 − 0.082 − 0.150 − 0.180 − 0.146 − 0.168
 (0.145) (0.141) (0.144) (0.142) (0.145) (0.142)
Third base (3B) 0.010 0.048 − 0.007 0.029 − 0.000 0.043
 (0.126) (0.122) (0.125) (0.122) (0.126) (0.122)
Catcher (C) 0.038 0.142 0.040 0.135 0.051 0.149
 (0.161) (0.156) (0.161) (0.156) (0.162) (0.156)
Center fielder (CF) − 0.141 − 0.105 − 0.176 − 0.161 − 0.165 − 0.141
 (0.127) (0.123) (0.127) (0.123) (0.127) (0.124)
Designated hitter 0.983**** 1.009**** 0.987**** 1.016**** 1.011**** 1.047****
 (DH)
 (0.225) (0.211) (0.222) (0.208) (0.222) (0.209)
Left fielder (LF) 0.076 0.046 0.061 0.023 0.068 0.049
 (0.134) (0.126) (0.135) (0.128) (0.135) (0.129)
Pitcher (P) − 1.291**** − 1.197**** − 1.317**** − 1.238**** − 1.304**** − 1.217****
 (0.124) (0.120) (0.124) (0.120) (0.124) (0.121)
Right fielder (RF) 0.264** 0.322*** 0.219* 0.273** 0.237* 0.299**
 (0.129) (0.125) (0.128) (0.124) (0.129) (0.126)
Constant − 4.483**** − 5.053**** − 5.054**** − 5.988**** − 5.212**** − 6.093****
 (1.344) (1.279) (1.330) (1.263) (1.331) (1.259)
Observations 1907 1907 1907 1907 1907 1907

Standard errors in parentheses using default VCE
*p < 0.10; **p < 0.05; ***p < 0.01; ****p < 0.001

count by any amount! Rather, only offensive WAR​(oWAR​) and pitching WAR​(pWAR​)
led to significant increases in MVP vote points before 2004. From 2004, improvements in
dWAR​ were rewarded meaningfully in the MVP race. As defense is the least salient source

 13
Public Choice

Table 3  Additional estimation (1) (2)
results
 VotePts VotePts

 oWAR​ 0.569****
 (0.02)
 oWAR_after 0.070****
 (0.015)
 dWAR​ 0.002
 (0.036)
 dWAR_after 0.104**
 (0.049)
 pWAR​ 0.721****
 (0.058)
 pWAR_after03 0.066**
 (0.032)
 Observations 1324 583

 Standard errors in parentheses using default VCE. Shortstop is the
 omitted position category
 *p < 0.10; **p < 0.05; ***p < 0.01; ****p < 0.001

of value on the baseball field, its lack of (statistical) importance in MVP voting before 2004
stands to reason. With the publication of dWAR​values, evidence exists that defensive value
became a productive input in the MVP race for the first time. We also find evidence that
voting from 2004 on rewarded marginal improvements in oWAR​and pWAR​more strongly.
All three types of win value in baseball were compensated more strongly by MVP voters
beginning in 2004.

3.1 Considering election “fairness” and relationships to political voting

The results reported in Table 3 suggest that voters are not only responsive to aggregate
information on candidate quality. They are even responsive to information that allows them
to grade candidates at the constituent source of value. That conclusion might be seen as
analogous to the candidate-issue level in political voting. Table 3 also suggests that vot-
ers can be less responsive or even unresponsive to less salient candidate attributes. How-
ever, voters also are shown to respond to mechanisms that bring improved clarity to such
attributes. In Table 4, we estimate other modified versions of model (3). The modified
models consider the potential effects of player age, race, and teammate productivity on
a player’s MVP vote count. We consider those variables while controlling for the player,
WAR​, whether or not the vote took place before 2004, and other characteristics from model
(3). In a meritorious MVP election, those additional variables would not influence the vote.
That is, they are each irrelevant to the issue of a player’s individual contribution to winning
baseball games. Hence, many of the results of Table 4 may be seen as a test of the “fair-
ness” or meritorious nature of MVP voting—before and after 2004—such that a fair vote
would be taken as one that ignores player characteristics that are irrelevant to the issue of
player on-field value.

13
Public Choice

Table 4  "Fairness" of voting (1) (2)
behavior estimation results
 VotePts VotePts

 WAR​ 0.535**** 0.521****
 (0.017) (0.020)
 war_after03 0.066**** 0.101****
 (0.012) (0.029)
 Age 0.058 0.064
 (0.082) (0.083)
 Age2 − 0.001 − 0.001
 (0.001) (0.001)
 teammate_war 0.027**** 0.028****
 (0.003) (0.003)
 teammate_war_after − 0.003
 (0.002)
 Asian 0.453* 0.584
 (0.275) (0.477)
 African American 0.081 0.102
 (0.079) (0.094)
 Hispanic 0.267**** 0.257***
 (0.065) (0.094)
 Asian after03 − 0.190
 (0.573)
 African American after03 − 0.075
 (0.149)
 Hispanic After03 0.017
 (0.130)
 Constant − 7.310**** − 7.392****
 (1.225) (1.244)
 Observations 1907 1907

 Standard errors in parentheses. Position variables specified but not
 listed for brevity
 *p < 0.10; **p < 0.05; ***p < 0.01; ****p < 0.001

 In the specifications of Table 4, we modify model (3) by replacing team fixed effects
with cumulative WAR​ of the player’s teammates in that season. Doing so, we find that a
unit increase in cumulative teammate WAR​ increases player vote points by approximately
2.7–2.8% points both before and after 2004, while the main result of interest (i.e., the coef-
ficient on WAR x after_2003) remains significant and positive in the modified model. That
result presents evidence that MVP candidates receive credit for teammate contributions and
that the credit was assigned both before and after 2004. That result is consistent with recent
evidence from a study of the NBA that finds teammate spillover effects in productivity
measures (Ghimire et al. 2020).
 Table 4 likewise reveals no evidence that player age influences MVP voting. Some evi-
dence that race can influence MVP vote points significantly exists even after controlling
for on-field player value. Hispanic players are estimated to receive more MVP vote points
than the reference group (i.e., non-Hispanic white players), ceteris paribus, at the = 0.05

 13
Public Choice

significance level. A Hispanic player earns an estimated 25.7% to 26.7% more vote points
than do other players, ceteris paribus. That bonus is equivalent to the estimated vote point
effect of a player gaining half a unit bump in WAR​. From the controls entered in Table 2′s
specifications, we find other results that potentially are relevant from the perspective of
political elections. Specifically, we find that democratic voting—in this case, equal voting
representation by MLB city—can overcome certain exposure advantages that may other-
wise accrue to “big city” or other high-profile candidates.
 On the basis of those results, we conclude that Baseball Writers are not simply writing
about advanced baseball statistics to add color to their articles. Since 2004, we find evi-
dence that the creation of WAR​ has changed voting behavior in a high-stakes environment
both significantly and substantially.

3.2 Robustness checks

We also assessed the robustness of our findings to alternative estimators, fixed effects, and
standard error corrections. We outline the robustness checks briefly here, but the results are
available by request. First, as discussed previously, we chose the negative binomial model
over the Poisson model because of concerns about overdispersion. However, an alterna-
tive approach is to estimate a fixed effects Poisson model, which will not suffer from those
problems as long as the standard errors are cluster-robust (Cameron and Trivedi 2009;
2013). We replicated our six models using that approach and found similar results. Next,
instead of specifying season fixed effects, we instead specified player-level fixed effects in
our six models. Again, we found very similar results. We also assessed whether the type of
standard error adjustment influenced our findings. We replicated our estimates using boot-
strapped standard errors, standard errors clustered at the season level, and standard errors
clustered at the player level. The results are robust to all of those alternative estimation
approaches.

4 Conclusion

This paper considers the effect of an exogenous change in the information frontier of voters
in a unique election setting, namely, the selection of the most valuable players (MVPs) in
the American and National leagues of Major League Baseball. We find that voters respond
to an instrument providing new information about candidate quality. Voting behavior is
more positively responsive to increases in candidate quality after the information fron-
tier shift. As a result, voting became less noisy overall with respect to observable candi-
date characteristics. Neutral measures of candidate quality (i.e., measures that reflect only
what a player does rather than who the player is) have ensured higher vote counts for more
qualified candidates, and the evidence suggests that the response is of a magnitude suf-
ficient to pivot elections. We observe that objective information about candidates is salient
to voters’ evaluations of candidates. In turn, voter behavior can change significantly and
substantially, whereby voters respond positively to objective information on stronger can-
didates as in Banerjee et al. (2011). We find significant and substantial evidence of such
responses not only with respect to the aggregate performance of candidates, but also with
respect to individual performance measures that matter to the electorate. Voters respond
to evidence of performance, where performance is found to be credited in each of its
measured forms. One particularly non-salient candidate characteristic—on-field defensive

13
Public Choice

performance—did not drive voting behavior whatsoever until voters were informed about
such performance. Thereafter, voters responded positively to evidence of superior defen-
sive skills. In that setting, voters do not exhibit a firmly fixed mindsets, whereby they might
be found to begrudge evidence that calls into question their past voting behavior. Rather,
members of the MVP voting body after 2004, many of whom also were members of the
same voting body beforehand, exhibited an average tendency toward behavioral agility in
processing new information about candidate quality. That is, they responded to the water-
shed introduction of the Wins Above Replacement (WAR) measure by which to compare
player win value cross-positionally.
 Our study offers several implications. First, it speaks to the debate on whether informed
voters choose higher quality governance or candidates (Banerjee et al. 2011; Pande 2011).
Our finding—that an exogenous information shock affects voter behavior—suggests that
they do. However, the results might also suggest that voters are conformists.14 That is, once
they accept that a computerized algorithm reflects the real qualities of candidate on the
ballot, they vote accordingly. If the algorithm changes, the voters may change their votes
in turn. Stated differently, voters can be manipulated easily. Future research might consider
the conditions under which information leads to voter manipulation, which has come to
light in recent years (Harvey 2016; Ziegler 2018). Second, it remains an open question
as to which types of elections our findings extend. Do our findings extend to jurors and
judges, whereby after receiving better information better decisions are made? Do our find-
ings extend to political elections, whereby voters cast votes based on a basket of issues?15
Ostensibly, our study extends more easily to scenarios like the latter, as WAR aggregates
win value from different sources (i.e., defense, offense, and pitching). The latter involves
ideology, multiple issues, and the risk of strategic voting and manipulation, which is more
complex. Nevertheless, the questions remain open and ripe for future research.
 Despite the availability of new information, some non-meritorious factors driving voter
behavior persisted even after 2004. Most prominently, player race and teammate produc-
tivity are found to influence voting behavior throughout the sample period. That finding
counters prior work on player race and elections to the Baseball Hall of Fame, which did
not document evidence of racial discrimination (Jewell 2003). Of course, MVP voting and
Hall of Fame voting are different. Future research might examine those and related issues,
especially concerning whether information affects discrimination in voting.

Acknowledgments We would like to thank William Shughart (and other members of the 2019 Public
Choice Society Meetings), Marek Kaminski, and anonymous referees for valuable feedback and suggestions.
For contributions during the early manuscript phase, we would also like to thank directors and judges of
the Carnegie Mellon Sport Analytics Research Competition, as well as Brian Taylor, Shana Gadarian,
Simon Weschle (and other faculty members and graduate students at the SU Maxwell School Department
of Political Science), Justin Ross, B. Andrew Chupp, and Seth Freedman (and other faculty members and
graduate students at the IU School of Public and Environmental Affairs). We would also like to thank Lara
J. Potter for her encouragement.

Funding This research was not funded.

Code availability Code is unpublished.

14
 We thank an anonymous reviewer for providing those thoughtful comments.
15
 Drazen and Eslava 2010 study voting scenarios based on a basket of issues.

 13
You can also read